Identification in Cascade Networks with Weighted Null-Space Fitting

• Main Author: Prota, Riccardo
• Format: Others
• Language: English
• Published: KTH, Skolan för elektroteknik och datavetenskap (EECS) 2018
• Subjects: Engineering and Technology; Teknik och teknologier
• Online Access: http://urn.kb.se/resolve?urn=urn:nbn:se:kth:diva-254395

System identification has the role to build mathematical models fordynamical systems, starting from the experimental data. One of themost studied and considered methods is the prediction error method(PEM); the main idea is minimizing a cost function based on the predictionerrors. The drawback with PEM is that it needs to solve a nonconvexoptimization problem. Weighted Null-Space Fitting (WNSF) isa three-step identification method which provides asymptotically efficientestimates without solving non-convex optimization problems.This could be an advantage in those cases where the system underconsideration is particularly complex.In the past decade, identification for dynamic networks has becomean important object of study, with applications in power systems, biologicalsystems, flexible mechanical structures, telecommunicationsystems and many others. The focus of this thesis is to extend theideas of WNSF to serial cascade networks. The objective is to estimatean entire dynamic network at once, without losing information andproviding asymptotic efficient estimates. For serial cascade networksthat satisfy certain properties in the location of inputs and outputs,we propose an identification procedure based on WNSF. We presentsimulations comparing performance with PEM, which highlight theadvantages of using WNSF as a method that does not have problemswith converging to non-global minima.For future work, we consider other network structures, which caninclude serial and parallel blocks. This work suggests that WNSF is anasymptotically efficient method for cascade networks and can be preferredor used as a complement to PEM when dealing with complexnetworks. === Syftet med systemidentifiering är att bygga matematiska modeller avdynamiska system från observerade data. Prediktionsfelsmetoden (PEM)är en standardmetod inom området, med asymptotiskt optimala egenskaper.En nackdel med denna metod är att den ofta kräver att en ickekonvexfunktion minimeras. Alternativa metoder som inte har sammanackdel har föreslagits, såsom viktad nollrumsanpassning (WNSF),som använder en viktad variant av minstakvadratmetoden och harsamma asymptotiskt egenskaper som PEM.Dagens alltmer komplexa system har gjort att intresset för dynamiskanätverk ökat. För nätverk blir PEMs konstadsfunktion ännu svårareatt minimera, vilket krävs alternativa algoritmer. I detta arbeteföreslår vi en metod för att applicera WNSF tpå kaskadnätverk. Metodensprestanda illustreras med simuleringar, vilka stöder hypotesenatt den är asymptotiskt effektiv.